A L U M I N I U M S M E LT I N G I N D U S T R Y

Testing a new ‘STARprobe’

M. Dupuis, GeniSim; J.-P. Gagné, Stas

In addition to the two main tasks an alumin- Taking into consideration the neutralisation means that if the direct AlF3 feed were to be
ium reduction cell controller has to perform, of some of the fluoride absorbed by the fresh completely stopped for some reason, it would
namely to keep both the dissolved alumina alumina in the scrubber by the sodium already take about 72 hours for the excess AlF3 con-
concentration in the bath and the anode cath- present in it, we can assume that the equiva- centration to drop by 1 to 9%. In view of this
ode distance (ACD) under tight control [1], lent of 3.6 kg/hr of AlF3 is fed back to the cell relatively slow response time of the cell, it
modern cell controllers are also in charge of by the secondary alumina (on average or at should be quite easy to keep the excess AlF3
keeping the bath ratio (or excess AlF3) concen- the nominal 100% alumina feeding rate). This concentration under tight control. But such
tration under control. leaves 1.1 kg/hr of AlF3 that must be directly control is clearly lacking in the great major-
This task has proven to be quite challenging fed using a point breaker feeder (PBF) under ity of smelters, so some other factors must be
despite the fact that, at first glance at least, it the supervision of the cell controller. complicating things.
looks quite straightforward. Fluoride evolves However, this hourly dose is only about How daily operations influence the bath
out of the cell in the off-gas; a big fraction of 0.14% of the excess AlF3 in the bath, since ratio: In the above mass balance calculation,
that fluoride is captured by the fresh alumina the cell contains close to eight tonnes of bath about 75% of the AlF3 is fed back to the cell
in the scrubber and returns to the cell as part of and hence about 800 kg of excess AlF3. This as part of the alumina feeding. However, in
the secondary alumina feed to it. The part that
does not return to the cell must be compen-
sated by direct AlF3 feeding in order to main-
tain a constant bath ratio in the cell. The cell
controller performs that task using feedback
control algorithms based on regular measure-
ments performed by cell operators.
Recently Alcoa has develop a revolutionary
new technology to measure bath ratio in the
potroom almost as quickly as you can meas-
ure bath temperature [2, 3]. Furthermore, in
addition to the excess AlF3 concentration, the
new STARprobe also measure the bath tem-
perature, the dissolved alumina concentration,
and the cell superheat. That last information
can be used as part of the cell control logic, as
previously presented in [4] for example.

GeniSim’s Dyna/Marc dynamic aluminium
reduction cell simulator has been used to mod-
el and evaluate the efficiency of the traditional
combined bath sample/XRD analysis and bath
© GeniSim
Fig. 1: Daily excess AlF3 concentration variation modelled without control and any mass imbalance as
generated by Dyna/Marc cell simulator

temperature measurement bath ratio control

logic, and to compare it with a new control
algorithm based on STARprobes excess AlF3
concentration and superheat measurements.

Performing the AlF3 mass balance

Using a 300 kA cell as an example, the fluo-

ride mass balance can be performed as fol-
lows. Fluoride evolved out of the cell at a rate
dictated by many factors, like the bath ratio
and temperature and the state of the anode
cover [5]. In the current example, the fluoride
evolution rate is calculated to be 33.6 kg F/t
Al with the cell conditions selected, namely
10% excess AlF3, 970 °C, and a good anode
cover. For a 300 kA cell producing 94.7 kg
Al / hr, this represents the equivalent of 4.7 kg
of AlF3 that evolves out of the cell and hence
must be replaced each hour. Fig. 2: 20 days excess AlF3 concentration variation modelled without control and any mass imbalance

76 ALUMINIUM · 1-2/2013
SPECIAL
Fig. 3: Corresponding 20 days of excess AlF3 concentration sampling results
modelled assuming no bath sampling noise
modern continuous tracking control logic, the
alumina is never fed constantly at the nomi-
nal 100% rate to the cell. As a result, the ex-
cess AlF3 concentration swings up and down
according to the alumina feeding cycle. The
direct AlF3 additions are also performed in
discrete events, for example 2 kg every 110
minutes, in order to average 1.1 kg/hr. Those
discrete additions also influence the short term
variation of the excess AlF3 concentration.
As well as the irregular AlF3 addition, sev-
eral thermal events also affect the AlF3 evolu-
tion, such as the bath temperature, but more
importantly the ledge thickness variation: as
ledge is mostly pure cryolite, ledge formation
concentrates the excess AlF3 in the molten
bath. Ledge formation occurs after anode
change events, for example. Fig. 1 shows the
calculated daily variation of the concentration
of AlF3 in the modeled bath in the absence of
Fig. 5: Simulation of the process without perturbation; top without control, bottom with feedback con-
trol, 10% target concentration (XRD results, once per day, 1 day delay, 0.5 kg/hr% proportional band and
-0.1 kg/hr°C proportional band)
ALUMINIUM · 1-2/2013
A L U M I N I U M S M E LT I N G I N D U S T R Y
Fig. 4: Corresponding 20 days of excess AlF3 concentration sampling results
modelled assuming 0.5% standard deviation white sampling noise
control additions and of any AlF3 mass imbal-
ance. The standard deviation on the average
value is about 0.1%.
Sampling frequency and delayed XRD re-
sults: The next factor complicating things is the
long delay in evaluating the bath chemistry
through manual interventions. The traditional
way of proceeding requires manual bath sam-
pling, manual processing of the bath samples,
at best semi-automatic analysis of the bath
samples by a XRD instrument, and manual
input of the results in a database accessible to
cell controllers. Considering the cost of a XRD
analysis, it is typical to take bath sample every
second day and to get results at 8 to 24 hours
after the actual bath sampling.
Fig. 2 shows the calculated variation of
the AlF3 for a period of 20 days, again in a
model without control additions or any mass
imbalance. Fig. 3 shows the results of the bath
sampling performed once a day, always at the
same time of the day. The delay between tak-
ing the sample and receiving the results of the
analysis is clearly not an issue when the con-
centration is drifting very slowly. Yet any delay
in the feedback response can cause instability,
depending on the controller setup.
So far, despite the daily events ‘process
noise’, the sparse sampling frequency, and the
delay in getting the sampling analysis results,
should make it easy to stabilise the bath ratio.
So there is no obvious explanation for why it
is so difficult to control the bath ratio.
Bath sampling noise problem: But a new
problem affecting bath ratio control has recent-
ly been identified: it is the bath sampling noise
due to the bath composition being far from ho-
mogeneous [6]. The standard deviation of that
bath sampling noise has been evaluated to be
around 0.5%, which is five times greater than
the process noise generated by daily events.
That bath sampling noise, contrary to the daily
events noise, is completely unpredictable. Fig.
4 shows the simulated results of bath sampling
performed on the 20 days period presented in
Fig. 2, but when 0.5% white noise is added to
the noise-free results presented in Fig. 3. The
fluctuations are about seven times greater.
Simulated process response using stand-
ard control without any process perturbation:
Now, we want to test a typical control logic
where both the delayed XRD analysis from
bath samples and measured bath temperature
are used to correct the direct AlF3 feeding
rate as it is commonly done in the industry
these days. The proportional band was set
to 0.5 kg/hr% for the 24 hours delayed bath
XRD analysis results, and to -0.1 kg/hr °C for
the bath temperature measurement. The bath
sampling and the temperature measurement
for the model are done simultaneously every
24 hours. ©
77
A L U M I N I U M S M E LT I N G I N D U S T R Y
A bath sampling noise having a standard de-
viation of 0.5% has been added to the XRD
analysis results following observation recently
reported [6]. For the temperature measure-
ment, a bath sampling noise having a stand-
ard deviation of 2.5 °C has been added as re-
ported in [6].
Fig. 5 presents the modeled results in the
dynamic cell simulator for a period of 100
days. The top graph presents the results ob-
tained without any control and in the absence
of process perturbation. The initial bump is
an indication that the steady state conditions
used as initial transient conditions are not
100% representative of the long term pseudo
steady state conditions.
The bottom graph presents the results ob-
Fig. 6: Simulation of 100 days natural response (no control) to a significant reduction of the anode cover
material thickness resulting in an increase of the anode panel heat loss by 30 kW
Fig. 7: Simulation of the process with a significant perturbation; feedback control, 10% target concentra-
tion (XRD results, once per day, 1 day delay, 0.5 kg/hr% proportional band and -0.1 kg/hr°C proportional
band)
78
tained with feedback control active. Unfortu-
nately, it is not as good as the results without
control. This shows that the bath sampling
noise, combined with the 24 delay in the bath
sampling analysis result, is destabilising this
feedback control loop.
Simulated process response using standard
control with a significant process perturba-
tion: In order to more seriously test the stabil-
ity of the feedback control loop, we added a
major perturbation to the simulation. On day
14, we simulated removal of about half of the
cover material from the anodes. This increas-
es the anode panel heat loss by about 30 kW
from 230 to 260 kW. As we can see in Fig. 6,
as a natural response, the cell must reduce its
cathode heat loss by the same amount. It does
this by reducing its superheat by about 1 °C
and by increasing its ledge thickness by about
5 cm. This extra ledge formation concentrates
the excess AlF3 in the molten bath by about
2%, where it remains close to 12% if the direct
AlF3 additions remain unchanged.
This is clearly a case where some feedback
control is required. Fig. 7 presents the model
results obtained using the standard control de-
scribed above. After the change of superheat,
the 970 °C temperature target is no longer
compatible with the 10% excess AlF3 target.
Combined with the 1 day offset between the
AlF3 feedback and the temperature feedback,
this generates a cyclic response characteristic
of somewhat unstable feedback control. This
type of oscillation with a wave length of about
20 days and an amplitude of about 2.5% is
very often seen in real smelters. Those unde-
sired oscillations occur despite careful selec-
tion of the values of the proportional constants
in an unsuccessful attempt to avoid feedback
loop instabilities.
The new STARprobe
The STARprobe is a portable device that takes
real-time measurements of bath properties
in electrolysis cells, such as Superheat, Tem-
perature, Alumina concentration and bath
Ratio or acidity (STAR). This synchronicity of
measurements is a most important step for-
ward in improving the control and efficiency
of electrolysis cells. It unites the conventional
processes of temperature measurement and
bath sampling analysis into one online meas-
urement. This simplifies and greatly shortens
the process and time delay from measure-
ment/sampling to pot control decision. The
pot control decision can therefore be based
on the real-time cell conditions rather than on
conditions from few hours ago, or even from
as long as 24 hours ago.
This integrated real-time measurement sys-
tem consists of four major components:
• Reusable probe tip
• Portable stand to fit various pot
configurations
• Electronics for data acquisition and
analysis, and wireless communications
for data transfer
• Tablet PC with programs to perform all
necessary tasks during measurements.
Considering the great advantages of the STAR-
probe, Alcoa has decided to share the technol-
ogy with the rest of the aluminium industry
starting from 2012. In this regard, Alcoa has
just appointed STAS, a well recognised leader
in the aluminium industry, to commercialise
the new STARprobe analysing system.
ALUMINIUM · 1-2/2013
SPECIAL
Simulated process response of STARprobe
to a significant process perturbation: Exactly
the same major perturbation was used to test
the efficiency of the feedback control loop
using STARprobe measurements. The same
day measurement frequency is used and the
same 0.5 kg/hr% proportional constant for the
AlF3 feedback loop. Obviously in this case
however, the measurement results are avail-
able without delay. In addition, the measured
superheat was also used to activate a separate
feedback loop which adjusts the target cell
resistance based on the offset between the tar-
get and the measured superheat.
The measured superheat is also affected
by a very significant bath sampling noise. That
bath sampling noise was estimated to have a
Fig. 8: Simulation of the process with a significant perturbation; feedback control, 10% target concentra-
tion (STARprobe measurements once per day, 0.5 kg/hr% proportional band and daily 0.1 µ: target resist-
ance correction due to superheat offset from target)
Fig. 9: Evolution of the cell target resistance (there is a 0.4 µ: change of target resistance each day during
the anode change event)
ALUMINIUM · 1-2/2013
A L U M I N I U M S M E LT I N G I N D U S T R Y
standard deviation of about 2 °C in [6], so we
added a 2 °C standard deviation white sam-
pling noise to the simulation.
The obtained results are shown in Fig. 8.
In this case the response to the perturbation
is slower than in the previous case, because
there is no longer a correction based on the
temperature offset, and because a ± 1 °C dead-
band was imposed on the superheat target in
order to inhibit wrong responses to the noise
in the superheat measurements. Yet after a 25
days transient response to the perturbation,
the excess AlF3 concentration goes back to
its target value and remains on target without
oscillations after that.
Fig. 9 shows the evolution of the target cell
resistance. After a delay of 6 days, a 0.01 μΩ
correction to the target cell resistance was
applied each day for 15 days giving a total
0.15 μΩ correction. This 0.15 μΩ ‘permanent’
correction ensures that the superheat remains
within the 3.5 to 5.5 °C range, despite the
fact that the anode panel now dissipates an
extra 30 kW of heat loss.
Conclusions
This study demonstrates the value of using a
dynamic cell simulator to optimise existing
cell controller algorithms and to test new ones
without putting real cells at risk. The Dyna/
Marc cell simulator used in this study is avail-
able to the whole aluminium industry through
GeniSim Inc. Version 14 supports adding the
simulated bath sampling noise at the level
seen in the AlF3 measurements. The model can
also use STARprobe measurements instead of
bath samples/XRD analysis to perform bath
ratio control.
The revolutionary new STARprobe meas-
urement tool makes possible a new control
logic scheme based on independent control
of the excess AlF3 and of the cell superheat.
Modelling proves this to be superior to the
standard single feedback control loop, which
uses two target variables (namely the excess
AlF3 and the operating temperature) to con-
trol a single control action, namely the direct
AlF3 additions.
The STARprobe developed by Alcoa [2,
3] is now available to the whole aluminium
industry through STAS (http://www.stas.com/
en/starprobetm.html).
References
[1] M. Dupuis, Testing cell controller algorithms
using a dynamic cell simulator, ALUMINIUM 88
(2012)1-2, 50-55.
[2] X. Wang, B. Hosler and G. Tarcy, Alcoa STAR-
probeTM, Light Metals, (2011), pp 483-489
[3] X. Wang, G. Tarcy, E. Batista, and G. Wood, Ac-
tive pot control using Alcoa STARprobeTM, Light
Metals, (2011), pp 491-496
[4] T. Rieck, M. Iffert, P.White, R. Rodrigo and R.
Kelchtermans, Increased Current Efficiency and Re-
duced Energy Consumption at the Trimet Smelter
Essen using 9 Box Matrix Control, Light Metals,
(2003), pp 449-456.
[5] W. Haupin and H. Kvande, Mathematical Model
of Fluoride Evolution from Hall-Héroult Cells,
Proceedings from the International Jomar Thons-
tad Symposium, ed. by A. Solheim and G. M. Haar-
berg, Trondheim, Norway, October 16-18, (2002),
53-65.
[6] M. Dupuis, P. Bouchard and J. P. Gagné, Meas-
uring bath properties using the STARprobeTM, 19th
International ICSOBA Symposium (2012), to be
published. ©
79
A L U M I N I U M S M E LT I N G I N D U S T R Y
Author
Dr. Marc Dupuis is a consultant specialised in the
applications of mathematical modelling for the
aluminium industry since 1994, the year when he
founded his own consulting company GeniSim Inc.
(www.genisim.com). Before that, he graduated with
Carbothermic reduction – An alternative aluminium production process
H. Kvande, NTNU
About twenty years ago the present author
co-authored a paper [1] with the title Car-
bothermal production of aluminium –
technically possible, but today economi-
cally impossible? Since then significant
resources have been spent on the study of
this process and much experimental work
has been done. So, it is time to ask again if
such a carbothermic process for production
of aluminium really is still economically
impossible. The present paper reviews the
published literature of the last two decades
to try to evaluate the current status of car-
bothermic aluminium production.
The standard industrial aluminium electroly-
sis process (Hall-Héroult) has several weak-
nesses, as we know: very high capital invest-
ment, a complex anode change operation,
high energy consumption, pollution of the en-
vironment and emissions of greenhouse gases.
That is why the search for alternative meth-
ods for production of aluminium will probably
never stop. In 2000 Alcoa announced that it
had started to develop a process based on car-
bothermic reduction of alumina. Since then
little published information has emerged
about the progress of Alcoa’s work. This is
quite understandable in view of the impor-
tance which a successful result would have.
So let us here first take a look at the present
status of carbothermic aluminium production.
Carbothermic production
of aluminium – its history
The idea of carbothermic reduction of alumi-
na to aluminium is indeed an old dream.
Aluminium-copper alloys with about 15% Al
were produced industrially already in 1886
[2], the same year as the present industrial
electrolysis process was invented. In the 1920s
Al-Si alloys with 40-60% Al were produced
in Germany, and about 10,000 tonnes of
these alloys were produced annually in the
period up to 1945.
80
a Ph.D. in chemical engineering from Laval Univer-
sity in Quebec City in 1984, and then worked ten
years as a research engineer for Alcan International.
His main research interests are the development of
mathematical models of the Hall-Héroult cell, deal-
ing with the thermo-electric, thermo-mechanic,
electro-magnetic and hydrodynamic aspects of the
The first attempt to produce pure aluminium
by carbothermic reduction of alumina was
performed around 1955. Pechiney worked
on the process from 1955 to 1967, but ter-
minated the programme for technical reasons.
Reynolds worked on an electric arc furnace to
produce aluminium from 1971 to 1984. Alcan
acquired information from Pechiney and con-
tinued their research, but stopped in the early
1980s. Alcoa tried to develop the process to
produce Al-Si alloys from 1977 to 1982.
However, in 1998 Alcoa started the carbo-
thermic production project again, together
with Elkem R&D in Norway. They changed
their focus from an open arc furnace (with high
generation of volatile aluminium-containing
gases) to a submerged arc process. Elkem
already had a long experience with modern
silicon furnace technology, and so came up
with the idea for a new type of high-tempera-
ture electric reduction reactor tailor-made for
carbothermic production of aluminium. Alcoa
had a good understanding of the fundamental
chemistry and a long experience with carbo-
thermic production of aluminium from the
work in the 1960s until the 1980s. Together
Alcoa and Elkem then agreed to try this again.
Carbothermic aluminium production:
the three main steps in the process
As the name says, the purpose of the carbo-
thermic method is to use carbon and heat to
reduce alumina to aluminium, according to
the overall reaction:
Al2O3(s) + 3 C(s) + heat = 2 Al(l) + 3 CO(g)
The reaction proceeds close to and above
2 000 °C, and it produces CO as the primary
gas. The gaseous by-product is therefore dif-
ferent from that of the Hall-Héroult process,
which produces CO2.
The carbothermic process can be divided
into three steps, as shown in the flow chart:
• Production of a slag, which contains a
molten mixture of alumina and alumin-
ium carbide
problem. He was also involved in the design of ex-
perimental high amperage cells, and in the retrofit
of many existing cell technologies.
Jean-Pierre Gagné is specialist for elecrolysis
products of Stas, based in Chicoutimi, Canada. Stas
designs and manufactures equipment for the pri-
mary and secondary aluminium industry.
• Production of a molten aluminium-
carbon-(carbide)-alloy
• Production of pure aluminium (refining)
from the aluminium-carbon-(carbide)-
containing alloy.
The two most difficult steps here are steps 2
and 3; the production of the molten alumin-
ium alloy and the subsequent refining of this
alloy. In addition the process needs a gas
scrubber to collect the aluminium-contain-
ing gases that evaporate from the furnace at
these high temperatures. This is an engineering
challenge. The main reactions are:
Overall carbothermic reduction:
Al2O3 (l) + 3 C (s) = 2 Al (l) + 3 CO (g) E°theoretical
= 7.9 kWh/kg Al
Stage 1 (T > 1 900 °C):
2 Al2O3 (s) + 9 C (s) => (Al4C3 + Al2O3) (slag)
+ 6 CO (g)
Stage 2 (T > 2 000 °C):
(Al4C3 + Al2O3) (slag) => (6 Al as metal alloy
with Al4C3) + 3 CO (g)
The latter two chemical equations are not
stoichiometrically correct here, because both
the slag and metal phases will have varying
compositions. The molten aluminium phase
will always contain some dissolved carbon,
and therefore it can be considered chemically
as an Al-C alloy. There are two molten phases
here and they will not mix. The molten alloy
has the lower density and will float on top of
the molten slag phase.
The carbothermic reduction process pro-
duces poisonous CO, which has to be captured.
If the CO were later burnt as fuel it would
produce CO2. To avoid releasing this green-
house gas this should then either be used as a
chemical or captured and stored (CCS).
Information published after year 2000
Here is a list of expected potential gains from
carbothermic aluminium production, as pub-
lished by Alcoa in 2000 [3]:
ALUMINIUM · 1-2/2013